Comparing Nvidia H200 with Chinese AI Chips: Performance Gap and Future Roadmap

Beijing:

US President Donald Trump has announced that the United States will permit the export of Nvidia's H200 processors to China, although it remains uncertain whether Beijing will authorize Chinese companies to purchase these chips.

China has developed several domestic AI chipmakers as part of its strategy to reduce dependency on foreign technology. Here's a comparison between Chinese competitors and Nvidia's H200.

The most advanced chip from Chinese manufacturers is Huawei's Ascend 910C, which still falls considerably behind the H200 in terms of computing power and memory bandwidth, according to recent research from the Institute for Progress.

Huawei's 910C delivers a total processing performance (TPP) of 12,032, compared to the H200's 15,840, and offers memory bandwidth of 3.2 terabytes per second versus the H200's 4.8 TB/s.

Other Chinese manufacturers such as Cambricon with its top chip, the Siyuan 590, and Hygon's BW1000 both lag behind Huawei's 910C in performance metrics.

When comparing with Nvidia's H20, a downgraded processor specifically designed for the Chinese market that ceased shipping earlier this year following Washington's ban, Chinese chips demonstrate competitive performance.

Huawei's 910B achieves a total processing performance of 5,120, surpassing the H20's 2,368, according to a July Bernstein report. Similarly, Cambricon's Siyuan 590, with a TPP of 4,493, also outperforms the H20.

Nvidia maintains its market dominance largely due to its CUDA software platform, which developers have extensively utilized for building AI models.

Transitioning to domestic chips would require developers to rewrite code and adapt to new platforms, representing a significant investment of time and resources.

Despite domestic chips demonstrating superior raw computing power compared to the H20, Chinese internet companies continue to prefer Nvidia's offerings due to the absence of a mature domestic alternative to the CUDA platform.

Huawei revealed its AI chip roadmap in September, announcing three new products scheduled for release over the next three years.

The Ascend 950PR is set to launch in Q1 2026, followed by a higher-memory variant, the 950DT, in Q4 2026. The Ascend 960 is planned for Q4 2027, with the Ascend 970 scheduled for Q4 2028.

According to a September Bernstein report, the 960's computing power approximately matches the H200. However, it features significantly enhanced interconnect bandwidth at 2,200 gigabytes per second compared to the H200's 900 GB/s.

Higher interconnect bandwidth facilitates faster communication between chips in multi-chip systems, which is essential for training large AI models. This suggests Huawei is prioritizing networking speed over raw computing capabilities.

The H200 represents an older generation of Nvidia's AI chips, built on the Hopper architecture launched in 2022, which predates the latest Blackwell architecture unveiled last year.

Nvidia's newest AI server, which integrates 72 Blackwell chips into a single computer, enhances the performance of certain AI models by 10 times compared to H200 servers, according to data recently published by the company.

A report released by the non-partisan think tank, the Institute for Progress (IFP), indicates that the Blackwell chip currently utilized by U.S. AI firms is approximately 1.5 times faster than H200 chips for training AI systems and five times faster for inferencing work, where AI models are operationalized.

Nevertheless, the H200 would be nearly six times more powerful than the H20, which is the most advanced AI semiconductor legally exportable to China, according to IFP's findings.